calendar detail

Speaker:
Charles Yang

A discovery procedure, as Chomsky famously discussed in Syntactic Structures, is a learning theory which, in combination of a universal theory of language, can produce an appropriate grammar on a corpus of language specific input. The pursuit of a discovery procedure was deemed premature. In one form or another, subsequent research in language acquisition and theoretical linguistics has set on the less ambitious project of evaluation procedure: Among a set of available grammars, which one is the best given a linguistic corpus?

I would like to suggest that a discovery procedure is now within striking distance. I will describe a principle of learning and generalization, which is almost surely domain general, that is capable of discovering systematic generalizations in a linguistic corpus but avoiding spurious ones. I will discuss its application to several long-standing puzzles in English morphology and syntax. Its mechanical nature will be illustrated by a live simulation of learning the infamous German noun plural system: first by me, someone who knows no German at all thereby simulating a German-learning child, followed by quantitative results from a fully implemented computational model.

A successful discovery procedure will eliminate much of the substance in Universal Grammar, while demarcating the boundary between what is linguistically explainable and what is not (or shouldn’t be).